Machine learning for pattern and waveform recognitions in terahertz image data.

Area of Science:

• Materials Science
• Spectroscopy
• Machine Learning

Background:

• Terahertz time-domain spectroscopy (THz-TDS) generates complex datasets.
• Automated analysis of THz-TDS data is crucial for applications like material characterization.
• Machine learning (ML) offers potential for pattern and waveform recognition in THz data.

Purpose of the Study:

• To evaluate the feasibility of various ML techniques for automated pattern and waveform recognition in THz-TDS datasets.
• To identify the most effective ML algorithm for analyzing THz spectral data, specifically for rust detection.

Main Methods:

• Several machine learning algorithms were tested on THz-TDS datasets.
• Random forest algorithm was specifically evaluated for its performance in both frequency and time domains.
• Classifier performance was assessed using out-of-bag error and correlation analysis.

Main Results:

• Random forest demonstrated superior performance for THz-TDS data analysis.
• A random forest classifier achieved less than 1% out-of-bag error for segmenting rusted and non-rusted regions in frequency-domain images.
• Linear correlation between rusted area, image resolution, and THz frequency served as a cross-validation metric.
• Standardized image pre-processing is required for consistent classifier application across different rust stages and limited to 1 ± 0.2 THz.
• Random forest excelled in classification accuracy and timing for time-domain waveform recognition.

Conclusions:

• Random forest is a highly effective ML algorithm for THz hyperspectral pattern recognition.
• The study validates the utility of random forest and other ML techniques for analyzing THz-TDS data.
• Further research may focus on optimizing pre-processing steps for broader applicability of ML classifiers in THz spectroscopy.